Emil Lynegaard

@tjruwase we ended up for the most part just using pure DDP in PyTorch. We did have moderate success using Fairscale which supported the variable batch sizes out of the...

Hey @tjruwase and @wangleiofficial As my original question is starting to be a bit old, I think I perhaps need to retest on a newer version of DeepSpeed to confirm...

Not an issue afaik. For about 12k instances, I experience similar evaluation times. If you want something faster (albeit results aren't 100% identical), I recommend [py-rouge](https://github.com/Diego999/py-rouge/issues)

@davidecaroselli any chance of including this in a future version?

In ModernMT casing is learned, as part of training, as there is no truecase usage or anything of the sort. So basically, this is just a result of your training....

@nicolabertoldi Hey Nicola, sorry for my slow response. I would like to test this out. Is this improved handling in the most recent release (v4.7) or do I have to...

@nicolabertoldi I just built v4.9.4 locally on my machine and tested it and this is still an issue. The queuing mechanism is not behaving as a FIFO queue.

This could be trivially added when in the future you update the version of `fairseq` as they have added native support for early stopping on the master branch (also means...

@nicolabertoldi doesn't seem like I can add the 'feature-requests' label myself, so feel free to add it.

As a temporary fix, I've simple written my config to look for the existence of a '.venv' folder, and use that as my jedi environment if it exists and otherwise,...